Light estimation from video is an interesting problem in many computer vision, computer graphics and robotics applications. For instance, in order to insert a virtual object into a video scene (a problem known as match moving or motion tracking in cinematography) and render it properly, the position of lights illuminating the scene should be known. Often light probes can be positioned at points within the video scene to measure the incident illumination conditions at each point within the scene. But using light probes can be time-consuming because such use requires physically placing a light probe at each point of interest within the scene and recording the illumination conditions at that point of interest. This is not possible for every situation. Light probes also require access to the scene. Such access may be challenging at a later date since the scene may have changed, the lights may have been moved, or the scene may have been destroyed.